StealthGPT Benchmark October 2026: Originality.ai 5.0 Results
Table of Contents
Originality.ai 5.0: the new hardest test
Full results across six detectors
Output quality: bypass without the word salad
New in the API: text preservation
How we tested
FAQ
Originality.ai released its 5.0 detection model on October 20, and it is the toughest detector we have ever tested. In our October benchmark, it flagged almost every humanized text we gave it. StealthGPT was the only humanizer that beat it with any regularity, passing 34% of the time. Undetectable AI passed 4% of the time, WriteHuman 6%, and the other eight tools did not pass once.
We ran 50 AI-written texts through 11 humanizers, then scored every output on six AI detectors and five quality measures. All of the data is public, so you can check every number in this post.
Key takeaways:
Originality.ai 5.0 changes the field. The old Originality model passed most humanizers. Version 5.0 passed almost none.
StealthGPT leads on 5.0 by a wide margin. 34% pass rate, more than five times the next best tool.
StealthGPT wins where detection is hardest. It has the top pass rate on the three hardest detectors (Pangram, GPTZero, and Originality 5.0) and the lowest average AI score across all six: 21.5%.
Quality held up. StealthGPT scored 88 out of 100 on writing quality, the best of any tool that also gets past detectors.
New: text preservation. API users can now lock links, HTML, headings, and other spans so they come back unchanged while the rest of the text is humanized.
Originality.ai 5.0: the new hardest test
scores.csv, StealthGPT October 2026 benchmark · 50 samples, pass = under 50% AI
Originality.ai launched version 5.0 on October 20, 2026, with humanizer detection as its headline improvement. Its announcement says 5.0 caught about 90% of humanized AI text in the company's own tests. Our results line up with that claim.
On the previous Originality model, most humanizers passed the majority of the time; Grubby and WriteHuman passed every sample. On 5.0, eight of the eleven tools passed zero samples. WriteHuman fell from 100% to 6%, and Undetectable AI from 98% to 4%.
StealthGPT passed 17 of 50 samples on 5.0, or 34%. It is the only tool with a double-digit pass rate. Originality now labels anything under 10% AI as likely original; by that stricter line, StealthGPT cleared 15 samples (30%) and no other tool cleared one.
A 34% pass rate is not a guarantee. Originality 5.0 still catches StealthGPT output about two times in three. But every other humanizer we tested is effectively blocked by it.
Full results across six detectors
StealthGPT posted the lowest average AI score of any humanizer, 21.5%, and the top pass rate on Pangram, GPTZero, and Originality 5.0. These are the three detectors that separate the field: on Pangram, the next best tool passed 40% of the time against StealthGPT's 88%.
Pass rate = share of 50 samples scored under 50% AI. Sorted by average AI score, lowest first.
| Humanizer | Pangram | GPTZero | Originality 5.0 | Originality (prev.) | ZeroGPT | Winston | Avg. Human Score |
|---|---|---|---|---|---|---|---|
| StealthGPT | 88% | 92% | 34% | 92% | 84% | 92% | 80.3% |
| Undetectable AI | 40% | 86% | 4% | 98% | 100% | 100% | 71.3% |
| WriteHuman | 26% | 90% | 6% | 100% | 96% | 100% | 69.7% |
| Rephrasy | 26% | 70% | 0% | 86% | 92% | 100% | 62.3% |
| Grubby | 10% | 68% | 0% | 100% | 96% | 100% | 62.3% |
| RewriteAI | 2% | 54% | 0% | 98% | 90% | 94% | 56.3% |
| Phrasly | 4% | 44% | 0% | 72% | 82% | 90% | 48.7% |
| Humbot | 0% | 0% | 0% | 76% | 88% | 100% | 44% |
| HIX Bypass | 0% | 6% | 0% | 66% | 92% | 100% | 44% |
| BypassGPT | 0% | 4% | 0% | 56% | 88% | 100% | 41.3% |
| Ryne | 0% | 0% | 0% | 0% | 4% | 0% | 0.7% |
| Unedited AI text | 0% | 0% | 0% | 0% | 2% | 0% | 0.3% |
On the easier detectors (ZeroGPT, Winston, and the previous Originality model), most humanizers pass nearly every time, and a few edge out StealthGPT by a handful of points. Those detectors no longer tell humanizers apart. The ones that do are Pangram, GPTZero, and now Originality 5.0, and StealthGPT leads all three.
Output quality: bypass without the word salad
StealthGPT scored 88 out of 100 on overall writing quality, the highest of any humanizer that also gets past detectors. It averaged 0.36 factual errors and 0.24 clunky sentences per sample, among the lowest in the test.
That matters because most humanizers trade quality for bypass. HIX Bypass, Humbot, and BypassGPT averaged more than four clunky sentences per sample and grammar scores around 70, and still failed Pangram and GPTZero almost every time.
| Humanizer | Quality (0-100) | Grammar (0-100) | Meaning kept (0-100) | Fact errors per sample | Clunky sentences per sample |
|---|---|---|---|---|---|
| Ryne* | 91.7 | 98.8 | 95.1 | 0.54 | 0.08 |
| StealthGPT | 88.0 | 96.6 | 94.8 | 0.36 | 0.24 |
| Grubby | 86.3 | 94.7 | 96.0 | 0.24 | 0.62 |
| Phrasly | 83.7 | 93.1 | 95.1 | 0.76 | 0.36 |
| RewriteAI | 77.0 | 97.0 | 89.3 | 1.04 | 0.84 |
| WriteHuman | 72.7 | 93.8 | 83.2 | 1.16 | 0.66 |
| Rephrasy | 71.0 | 89.7 | 83.3 | 1.38 | 1.44 |
| Humbot | 68.0 | 70.5 | 81.7 | 1.30 | 4.22 |
| Undetectable AI | 65.3 | 91.1 | 85.7 | 1.44 | 1.36 |
| HIX Bypass | 65.3 | 70.1 | 80.3 | 1.60 | 4.24 |
| BypassGPT | 65.0 | 68.6 | 80.8 | 1.58 | 5.00 |
*Ryne scored highest on quality because it barely changes the input. Its output averaged 96% AI across six detectors, almost identical to unedited AI text.
The two tools closest to StealthGPT on Originality 5.0 paid for it in quality. Undetectable AI scored 65.3 with 1.44 factual errors per sample; WriteHuman scored 72.7 and kept the least of the original meaning among the top three.
New in the API: text preservation
StealthGPT now lets you lock parts of your text so they come back exactly as you wrote them, while everything around them is humanized. Links, HTML tags, headings, product names, citations, and code snippets stay intact, byte for byte.
This fixes one of the most common problems with humanizers. Run a blog post with links and headings through most tools and you get rewritten anchor text, broken HTML, and reworded brand names. With text preservation, you swap each span you want to protect for a short marker, send the text, and swap the markers back.
{
"prompt": "Read [[A1]]our pricing guide[[A2]] before you buy.",
"rephrase": true,
"model": "super",
"preservation": true
}
How it works:
Marked spans never change. Every marker comes back in the same order, copied from your input rather than the model. If any marker can't be returned unchanged, the request fails instead of returning broken text, and you aren't charged.
Bypass stays strong with Super. With the Super model, markers never go to the model. The full text is rewritten in one pass and placed back around the markers sentence by sentence, which keeps most of Super's detection bypass.
Preserved words are free. Words inside markers aren't billed.
Custom delimiters. Use
[[ ]]by default, or set your own opening and closing markers if your content already uses double brackets.
For best results, put markers for headings and other blocks on their own line or between sentences. Text inside markers isn't humanized, so the more you freeze, the more detectable the remaining text can be. Full details are in the text preservation docs.
How we tested
Samples. 50 AI-generated texts, 364 to 997 words each (669 on average). Before humanizing, the unedited text averaged 97.2% AI, and only one sample passed a single detector (ZeroGPT).
Humanizers. Each sample ran through 11 tools: StealthGPT, Undetectable AI, WriteHuman, Rephrasy, Grubby, RewriteAI, Phrasly, Humbot, HIX Bypass, BypassGPT, and Ryne. We used each tool's default settings.
Detectors. Every output was scored on Pangram, GPTZero, Originality.ai 5.0, the previous Originality.ai model, ZeroGPT, and Winston AI.
Pass rate. A sample passes a detector when it scores under 50% AI. Average AI score is the mean of all six detectors.
Quality. Each output was scored for overall quality, grammar, and how much of the original meaning it kept, plus counts of factual errors and clunky sentences.
We built and ran this benchmark ourselves, and we publish every score so anyone can check it. The full per-sample data is in our public benchmark data on GitHub.
FAQ
Can you bypass Originality.ai 5.0?
Yes, and right now only with StealthGPT. In our October 2026 test, StealthGPT passed Originality 5.0 on 34% of samples. Undetectable AI passed 4%, WriteHuman 6%, and the other eight humanizers we tested never passed.
What changed in Originality.ai 5.0?
Originality.ai built 5.0 to catch humanized AI text, and released it on October 20, 2026. It also added a slower, more expensive Max mode for high-stakes checks. Humanizers that passed the old model almost every time now fail it almost every time.
What is the best AI humanizer in October 2026?
In our benchmark, StealthGPT had the lowest average AI score across six detectors (21.5%) and the top pass rate on Pangram (88%), GPTZero (92%), and Originality 5.0 (34%). It also had the best writing quality of any tool that gets past detectors.
Does humanizing hurt writing quality?
With many tools, yes. Three of the eleven tools averaged four or more clunky sentences per sample. StealthGPT averaged 0.24 clunky sentences and 0.36 factual errors per sample, with a quality score of 88 out of 100.
Can I humanize text without breaking links or formatting?
Yes. StealthGPT's text preservation feature, available in the API, locks links, HTML, headings, and any other marked spans so they come back unchanged. Words inside preserved spans aren't billed.
Is the benchmark data public?
Yes. Every per-sample score from this test is published on GitHub, so you can verify any number in this post.
